Online non-affine nonlinear system identification based on state-space neuro-fuzzy models

This paper proposes a new general recurrent state-space neuro-fuzzy model structure. Three topologies are under assessment, including the state-input recurrent neuro-fuzzy system, the series-parallel recurrent neuro-fuzzy system and the parallel recurrent neuro-fuzzy system. Moreover, the underlying generalised state-space Takagi–Sugeno system is proven to be a universal approximator, and some stability conditions derived for this system. The online training is carried out based on a constrained unscented Kalman filter, where weights, membership functions and consequents are recursively updated. Results from experiments on a benchmark MIMO system demonstrate the applicability and flexibility of the proposed system identification approach.